How can we make sure a prompt stays effective when translated into another language? What techniques should we use for this? Who has faced these kinds of problems before?
1 Like
Depending on the use case and the language, you may want to do all the instruction in English and only translate the final output to the language of the user.
For other use-case, like RAG with multi-language documents, you may want to translate a user query or expand it in each of the languages.
Finally, if you only need to support a few languages, you can write the necessary prompt chains in each language and use a language detector (simpler classifier or LLM if you want) to use the right-language chain.
Dynamic translation of system prompts can render very unpredictable results, Iād generally not recommend it.
Good luck!
4 Likes
Agree with this 100%
Also, I am seeing more AI models coming out already built on understanding local languages. You could try those too.
2 Likes